Real-Time Fair-Exposure Ad Allocation for SMBs and Underserved Creators via Contextual Bandits-with-Knapsacks

In the realm of digital advertising, allocating exposure resources among small to medium-sized businesses (SMBs) and disadvantaged content creators in an efficient and fair manner is a significant and challenging issue. This paper introduces FairCBwK (FCBwK), a joint optimization framework for real-time fair exposure advertising based on the "Contextual Bandits with Knapsacks (CBwK)" paradigm. The proposed approach incorporates Fairness-of-Exposure constraints at both group and individual levels, alongside traditional revenue metrics such as maximizing click-through rate (CTR), AUC, or calibration accuracy. This aims to minimize the disparity in impression share between advertisers while ensuring both interpretability and fairness in exposure allocation. In terms of algorithm design, FairCBwK builds upon existing Fair-CBwK literature, pragmatically combining reward-budget dual optimization strategies in reinforcement learning with decoupled exposure modules, and dynamically adapting coupling weights between revenue, fairness, and budget using a Lagrangian multiplier optimizer. Experimental results demonstrate that FairCBwK significantly reduces exposure imbalances while achieving high CTR and budget utilization rates.